A High Accuracy Lifetime Estimation Method for SiC MOSFETs Based on CNN-LSTM Networks

Abstract

Accurate lifetime prediction of power devices is critical for preventing unexpected converter failures and extending service life. However, existing methods often suffer from limited prediction accuracy, reliance on extensive experiments, and difficulties in practical implementation. This paper proposes a data-driven lifetime estimation method for silicon carbide (SiC) MOSFETs based on the prediction of a single aging precursor, i.e., the on-state drain-source voltage drop ( $V_{\mathrm {ds\_on}}$ ). The proposed method integrates the local feature extraction capability of convolutional neural networks (CNNs) with the temporal modeling strength of long short-term memory (LSTM) networks, enabling effective capture of complex degradation trends and improved prediction accuracy. Moreover, it facilitates practical implementation, as $V_{\mathrm {ds\_on}}$ is easy to measure online during normal operation. A DC power cycling test setup with online $V_{\mathrm {ds\_on}}$ sampling is designed to accelerate device aging and collect degradation datasets. Multiple devices can be aged simultaneously under different operating conditions using the proposed setup, improving test efficiency. Experimental results under various aging conditions demonstrate that the proposed method achieves the highest prediction accuracy of 99.3%, outperforming the LSTM-based and exponential model-based approaches, which achieve average accuracies of 96.4% and 41.1%, respectively. These advancements are beneficial for enhancing system reliability and reducing operational costs in practical applications.